1. Introduction
Categorization involves a series of basic cognitive operations that allow organizing the information we receive from a complex and multifaceted reality. Their study helps us to understand how human beings interact with the environment and how we structure knowledge (Harnad, Reference Harnad, Cohen and Lefebvre2005). According to linguistic theory, categorizing processes also determine the nature of the language we use to communicate and create a socially shared world (Kleiber, Reference Kleiber1990). Western linguistic tradition has provided several models of categories for potentially grouping the concepts that have orchestrated knowledge since Aristotelian times (Boys-Stones, Reference Boys-Stones and Boys-Stones2017) – enriched by the contributions stemming from cognitive psychology (Barsalou, Reference Barsalou1983; Rosch, Reference Rosch1975) and cognitive linguistics (Lakoff, Reference Lakoff1987).
From a different standpoint, in psychology, categorization in verbal behaviour has been studied by examining responses of speakers to various semantic categories within an extensive array of tasks. Since the mid-20th century, and particularly after the influential seminal work of Battig and Montague (Reference Battig and Montague1969), whose methodology has been extensively replicated, research on categorial association has aimed to understand how individuals assign exemplars to specific categories. These authors collected normative data on responses to 56 verbal categories from English speakers through a category fluency task within 30 seconds (e.g., ‘List all the animals names that come to mind’ or ‘Write as many examples as you can for the category Musical instruments’). They explored the linguistic outcomes across a wide range of dimensions, analyzing measures of exemplars – the words provided by participants for each category label – such as the frequency of unique exemplars in each category and their average output position, and category-level measures like the potency of exemplar generation of a specific category label or categorial stimulus (‘Food’, ‘Tools’, ‘Colours’), among others. Consequently, psycholinguistic categorization studies have highlighted various levels of association within the lexicon of natural languages: between the categorial stimulus and its exemplars, and among responses within a category, revealing specific underlying mechanisms in the mental lexicon (Aitchison, Reference Aitchison2012).
A recent paradigm for a further exploration of relationships as part of our mental lexicon and, in general, the structure of knowledge, involves graph theory and the construction of complex networks. The mental lexicon is not a static reality but instead a dynamic cognitive system that constitutes the capacity for conscious and unconscious lexical activity (Jarema & Libben, Reference Jarema and Libben2007). The architecture of the mental lexicon is better understood as resulting from the dynamic integration of multiple levels of information (Pirrelli et al., Reference Pirrelli, Plag and Dressler2020). Lexical-semantic networks may harness this dynamism, thereby making them a suitable model for mental lexicon analysis.
Steyvers and Tenenbaum (Reference Steyvers and Tenenbaum2005, p. 72) define a lexical-semantic network as ‘an abstract representation of one aspect of semantic knowledge – the pairwise relations between words’. The creation of a lexical-semantic network based on experimental linguistic data does not mean creating models of the human mind and knowledge, but rather abstract and metaphorical representations of a type of relations between the words present in knowledge (Aitchison, Reference Aitchison2012; Fitzpatrick & Thwaites, Reference Fitzpatrick and Thwaites2020). Although a network based on experimental or behavioural data is a simplification of the mental lexicon’s structure, linguistic networks provide a privileged internal perspective for its description that could not otherwise be observed. A network representation always captures the interplay of language and cognition (Beckage & Colunga, Reference Beckage, Colunga, Mehler, Lücking, Banisch, Blanchard and Job2016), given that ‘structure always affects function’ (Strogatz, Reference Strogatz2001, p. 268).
The creation of complex networks has involved different kinds of linguistic data. Although networks have been formed with phonetic-morphological or syntactical data, the level of lexical analysis is undoubtedly the one that has been most widely studied according to network theory (Mehler et al., Reference Mehler, Lücking, Banisch, Blanchard and Job2016). Nevertheless, the type of lexical data included in the networks in different studies is of a highly varying nature (see the comparisons between them in Steyvers & Tenenbaum, Reference Steyvers and Tenenbaum2005; Borge-Holthoefer & Arenas, Reference Borge-Holthoefer and Arenas2009, Reference Borge-Holthoefer and Arenas2010; De Deyne et al., Reference De Deyne, Kenett, Anaki, Faust, Navarro and Jones2017). Besides computer-simulated networks (Steyvers & Tenenbaum, Reference Steyvers and Tenenbaum2005), there are studies that incorporate data based on textual corpuses that analyse the co-occurrence of words (De Deyne et al., Reference De Deyne, Kenett, Anaki, Faust, Navarro and Jones2017); their sources are dictionaries and lexicographic tools, such as WordNet (Steyvers & Tenenbaum, Reference Steyvers and Tenenbaum2005), while others are based on the analysis of semantic features breaking the words’ meaning down into their component parts and linking them depending on whether or not they share those features (McRae, Reference McRae and Ross2004) or articulating their sensorial features according to embodiment theory (Alvarado et al., Reference Alvarado, Velasco and Salgado2024). Finally, there are also networks created with experimental tasks involving verbal fluency that are designed to convey real empirical data to the model of graph theory.
There are two main paradigms of verbal fluency for extracting this type of data: continuous free word association (De Deyne & Storms, Reference De Deyne and Storms2008) and category fluency tasks – also called semantic fluency or exemplar generation – (Cosgrove et al., Reference Cosgrove, Beaty, Diaz and Kenett2023; Mazzuca & Majid, Reference Mazzuca and Majid2023). In the first task, participants generate words freely associated with a given term within a set time or response limit, resulting in broad associations (e.g., stimulus: sun; responses: moon, night, sleep, dreams, success, while). In contrast, the second task requires responses to be exemplars of a specific semantic category, leading to more structured categorial associations (e.g., stimulus: animals; responses: dog, cat, horse, sheep, goat). Both are particularly pertinent for describing language and cognition; the first because association is one of the principles of the organization of memory, and the second because taxonomical organization is one of the core processes of harnessing experience. Both, therefore, explain how to extract information from the semantic memory – the cognitive system that stores and allows extracting knowledge on words – although they make the relations of certain lexical units prevail over others. There is a very high presence of different kinds of associative relations and concurrence in free association tasks. Taxonomic relations (hyponyms, hypernyms and base terms) abound in category fluency; yet, as we shall see, there are also extensive thematic relations (De Deyne et al., Reference De Deyne, Kenett, Anaki, Faust, Navarro and Jones2017).
The previous literature identifies three structural levels for studying networks derived from linguistic data: the macroscopic level, which provides a global and holistic description of the structural properties of a network; the mesoscopic level, which examines the topology of communities or subgroups identified within the network and the microscopic level, which focuses on the context of specific nodes (Borge-Holthoefer & Arenas, Reference Borge-Holthoefer and Arenas2010; De Deyne et al., Reference De Deyne, Kenett, Anaki, Faust, Navarro and Jones2017; Steyvers & Tenenbaum, Reference Steyvers and Tenenbaum2005). The mesoscopic and microscopic approaches are the most suitable ones for analysing the internal structure of conceptual categories and the processes of semantic categorization, as the former permits analysing the fragment within the network’s macrostructure consisting of the exemplars contained in that category, and the latter identifies terms or clusters thereof that play a significant role in creating the structure.
This study’s overriding purpose, therefore, is to analyse the networks derived from the data obtained in a semantic fluency task for the following four highly significant models of conceptual categories within linguistic tradition – natural or taxonomic (Rosch, Reference Rosch1975), ad hoc (Barsalou, Reference Barsalou1983), radial (Lakoff, Reference Lakoff1987) and schema (Mandler & Cánovas, Reference Mandler and Cánovas2014) – with a view to verifying whether the networks generated confirm their criterion of differential linguistic description.
Mathematical tools from complex network theory will be used to describe network dynamics and structure. Besides traditional measures in the exploration of networks, we propose additional more informative ones created for this research. The main methodological innovation is that the measures used will not only provide us with information on the network structure but also, and particularly, on how speakers behave within it when they access linguistic data. This requires exploring the network not just as a construct but also as a process in whose development speakers are crucially involved.
The following are, therefore, our research questions:
-
1. Do network structure and dynamics reflect the differences in the type of conceptual categories within the semantic memory?
-
2. On a mesoscopic level: are there differences in the structure of the network generated through categories of a different nature according to the linguistic description? Are there differences in the process that speakers use to negotiate the network?
-
3. On a microscopic level: do the main terms in the networks for each one of the categories generate the same kind of structural dependence around them?
2. Theoretical framework
2.1. Semantic networks in the mental lexicon
The aim of studying the mental lexicon is to understand the knowledge we gain on words throughout our lives. The model of lexical networks has been widely developed according to the precepts of network theory and it represents the relations established between words that make up the lexicon within the semantic memory (De Deyne et al., Reference De Deyne, Kenett, Anaki, Faust, Navarro and Jones2017). The purpose of far-reaching networks based on data of continuous free association (De Deyne & Storms, Reference De Deyne and Storms2008) or computer simulations (Borge-Holthoefer & Arenas, Reference Borge-Holthoefer and Arenas2010) is to create simulations that are as similar as possible to the overall lexicon. This approach is informed by the seminal study by Collins and Loftus (Reference Collins and Loftus1975), who contend that a network is a valid model of psychological representation for describing the mental lexicon; that is, a metaphorical representation that reflects the structure of the latent lexicon, given that it explains its properties, such as the spread of activation, the extraction of meaning through the network and the growth of the network through learning. The words in this large-scale overall network establish relations based on different kinds of properties: semantic (dog-cat); syntactical or combinatory (eat-cake); phonological and orthographic (trick-truck), pragmatic or of usage (dear-sir) (Aitchison, Reference Aitchison2012), which validate a representation of experiential semantic knowledge, a sensory-motor interaction-based one and a language-based one (Vigliocco et al., Reference Vigliocco, Meteyard, Andrews and Kousta2009).
The types of relations between network nodes have prompted some scholars to differentiate between lexical networks and semantic (or thematic) ones. The former incorporate all nature of relations between nodes and tend to be created with continuous free association data, with contextually guided relations prevailing, whereas the latter are based on hierarchical taxonomic relations. Nevertheless, the boundaries between these two types of networks are blurred, as the responses the speakers provide in the fluency tests often interfere with all the processes involved in the meaning, understood in a broad sense, which incorporate the denotative meaning (semantic memory) and combinatory or experiential processes (associations) (Sánchez-Saus Laserna, Reference Sánchez-Saus Laserna2019).
Some of the more significant processes in the organization of knowledge and the systemization of experience involve categorization, being reflected in the semantic memory’s arrangement of the lexicon in the form of categories or word groups with a similar semantic or existential content. This categorization of the lexicon permits accessing the subnetworks or partial networks whose main reason for association involves relations of meaning. This level of analysis has been referred to as a mesoscopic description of the networks (Borge-Holthoefer & Arenas, Reference Borge-Holthoefer and Arenas2010). This approach analyses partial sections of a larger structure, which in network theory are called modules or communities, and provide information on the content, qualitative properties and relations of similarity within the lexicon (De Deyne et al., Reference De Deyne, Kenett, Anaki, Faust, Navarro and Jones2017). Linguistic categories could be included within this level of analysis by identifying regions within the overall network with a certain degree of meaning-based clustering. These regions tend to be hierarchical and allow identifying well-defined subcategories or communities within them.
The paradigm of category fluency serves as a robust empirical method for investigating semantic representations and the organization of conceptual storage (Goñi et al., Reference Goñi, Arrondo, Sepulcre, Martincorena, de Mendizábal, Corominas-Murtra, Bejarano, Ardanza-Trevijano, Peraita, Wall and Villoslada2011). Recent research has emphasized the development of standardized practices for constructing networks derived from category fluency data (Christensen & Kenett, Reference Christensen and Kenett2023; Zemla & Austerweil, Reference Zemla and Austerweil2018), driven by the growing application of this methodology in diverse linguistic inquiries. Empirical findings obtained through the analysis of network structures have contributed to understanding the relation between lexical-semantic knowledge and biological variables, such as aging (Wulff et al., Reference Wulff, Hills and Mata2022) and psychological features, as creativity (Kenett & Faust, Reference Kenett and Faust2019). Additionally, this paradigm has been instrumental in examining the lexicon structure of second-language speakers, facilitating comparisons with native speakers (Feng & Liu, Reference Feng and Liu2024; Hernández Muñoz et al., Reference Hernández Muñoz, Tomé, López García, Díaz and Muñoz2025; Li et al., Reference Li, Jiang, Shang and Chen2021). For example, research indicates that native speaker networks exhibit lower modularity than those of non-native speakers, thereby enhancing connectivity within the global network (Borodkin et al., Reference Borodkin, Kenett, Faust and Mashal2016). Moreover, it has been utilized to evaluate the impact of various factors on language acquisition, such as immersion experiences (Chaouch-Orozco & Martín-Villena, Reference Chaouch-Orozco and Martín-Villena2024), or increased educational exposure and enhanced proficiency levels (Agustín-Llach, Reference Agustín-Llach2022; Quintanilla Espinoza & Kloss Medina, Reference Quintanilla Espinoza and Kloss Medina2024). Category fluency not only provides access to network structures at macro and mesoscopic levels but also allows for microscopic analyses. So, this enables the identification of specific nodes or clusters that establish highly cohesive substructures in learning new-words procedures, which in turn contribute to the network’s overall integrity (Agustín-Llach & Rubio, Reference Agustín-Llach and Rubio2024). Finally, another field of research, where the analysis of semantic network structures is contributing noteworthy findings is the assessment of the linguistic performance associated with pathological conditions, including Alzheimer’s disease (Bertola et al., Reference Bertola, Mota, Copelli, Rivero, Diniz, Romano-Silva, Ribeiro and Malloy-Diniz2014; Lerner et al., Reference Lerner, Ogrocki and Thomas2009).
The large-scale lexical networks that represent the mental lexicon have two constant structural features: they are small-world networks and have a certain degree of modularity (Borge-Holthoefer & Arenas, Reference Borge-Holthoefer and Arenas2010; Steyvers & Tenenbaum, Reference Steyvers and Tenenbaum2005). Networks with a small-world structure have a high degree of clustering and a short overall distance, which means passing from one node to another in only a few steps (as opposed to random networks in which there is little local clustering and the overall distance is short), whereby information flows rapidly and efficiently without the drawbacks of hyperconnectivity The usual way of deciding whether a network is a small-world one involves considering two structural properties, namely, the clustering coefficient (CC) and the average shortest path length (ASPL). The distance between one node and all the others tends to be short in small-world networks. Furthermore, the degree of small-worldness is established through the S metric (Humphries & Gurney, Reference Humphries and Gurney2008), which requires an equilibrium between CC and ASPL (S values higher than 1 indicate a small-world network). A network’s degree of modularity refers to its division into communities or subcategories. A network is modular when it contains a dense cluster of communities (Newman, Reference Newman2010)). While a modular network tends to be a small-world one, not all small-work networks are necessarily modular (Borodkin et al., Reference Borodkin, Kenett, Faust and Mashal2016). These basic features for structuring the network generated by the categories constitute our starting point for confirming that our data replicate prior lexical-semantic models, although differentiating conceptual categories requires a more sophisticated analysis model.
In addition to the network’s topology, it is essential to address the dynamic processes that can be observed within it. Dynamic systems within network theory refer to networks that evolve over time according to specific rules or equations (Borge-Holthoefer & Arenas, Reference Borge-Holthoefer and Arenas2010; Newman, Reference Newman2010). Consequently, explaining the dynamics of a network involves understanding the processes driving structural changes, how a particular configuration emerges and how it evolves across time. As highlighted by Siew et al. (Reference Siew, Wulff, Beckage and Kenett2019), semantic and lexical networks exhibit inherent dynamism in various ways. On one hand, linguistic input and experiences enhance lexical knowledge throughout an individual’s lifespan as part of language development. Therefore, examining changes in networks during native and non-native language acquisition and learning processes becomes essential (Agustín-Llach, Reference Agustín-Llach2022; Chaouch-Orozco & Martín-Villena, Reference Chaouch-Orozco and Martín-Villena2024). On the other hand, lexical networks also exhibit dynamic phenomena, such as lexical access processes, the spread of activation between words, memory retrieval or the identification of central nodes that individuals apparently use to navigate the network (Baronchelli et al., Reference Baronchelli, Ferrer-i-Cancho, Pastor-Satorras, Chater and Christiansen2013; Siew et al., Reference Siew, Wulff, Beckage and Kenett2019).
Our study investigates the dynamics of categoril networks in the latter sense, focusing specifically on how speakers navigate these networks – namely, the processes of selecting and retrieving lexical items during semantic fluency tasks. In this context, the network is constructed by the speaker through the retrieval of semantic memory. Thus, this research adopts a more speaker-centred perspective on network dynamics, emphasizing not only the propagation of lexical information between nodes in an established network but also the preceding processes by which the speaker generates the network’s structure. This perspective aligns with a linguistics approach grounded in speaker-based analysis.
2.2. Categorial structures
In the framework of cognitive sciences, categorization is conceived as a mental process of organizing the information obtained through the apprehension of reality, which allows us to make sense of experience. ‘Categories are kinds, and categorization occurs when the same output occurs with the same kind of input, rather than the same input’ (Harnad, Reference Harnad, Cohen and Lefebvre2005, p. 22). This process, ‘by which distinct entities are treated as equivalent, is one of the most fundamental and pervasive cognitive activities’ (Wilson & Keil, Reference Wilson and Keil2001, p. 104), hence the great interest it has generated across fields like psychology, anthropology, philosophy, AI or neuroscience, in addition to linguistics (see Aarts, Reference Aarts2006 for a review in the history of linguistics). Even very young children, who are just beginning to produce their first words, are able to use language to categorise the elements around them. The earliest words designate members of well-organized categories (Koenig & Woodward, Reference Koenig, Woodward and Gaskell2007, p. 621) and countable nouns, in particular, seem to encourage category formation and recognition of patterns (Balaban & Waxman, Reference Balaban and Waxman1997; Graham et al., Reference Graham, Kilbreath and Welder2004; Pomiechowska & Gliga, Reference Pomiechowska and Gliga2019). Neuroscience further reveals that this phenomenon has a physical correlate. Different categories have distinct neuroanatomical bases (Damasio et al., Reference Damasio, Grabowski, Tranel, Hichwa and Damasio1996; Ullman, Reference Ullman and Gaskell2007): ‘the same regions are active, at least in part, when objects from a category are recognised, named, imagined, and when reading and answering questions about them’ (Martin & Chao, Reference Martin and Chao2001, p. 199).
Categorization is such a complex phenomenon that no model has been able to provide a satisfactory answer that encompasses all categories of human knowledge (Wilson & Keil, Reference Wilson and Keil2001, p. 104). For this reason, our verbal fluency task employs four categories that fit the different structural descriptions postulated by some of the main theories of categorization, namely, the prototype theory, the exemplar view and the knowledge approach (see Taylor, Reference Taylor2005). Firstly, Animals is considered a natural category, containing exemplars that reveal a relation of existential belonging to the group (a dog is an animal) and is organized around a central prototype, understood to be a combination of typical properties according to the standard version of prototype theory (Kleiber, Reference Kleiber1990; Rosch et al., Reference Rosch, Mervis, Gray, Johnson and Boyes-Braem1976; Rosch & Mervis, Reference Rosch and Mervis1975). As opposed to the classic Aristotelian concept, these properties are not the category’s defining features (in the sense of the necessary and sufficient conditions for its definition), and not all the members have to share them. By contrast, the relation that links them is family similarity, a structure for linking the exemplars in a category without having to share a common attribute that defines it (Lewandowska-Tomaszczyk, Reference Lewandowska-Tomaszczyk, Geeraerts and Cuyckens2007, p. 146).
In turn, the knowledge approach assumes that concepts are represented and organized according to people’s world views (Berko & Bernstein, Reference Berko and Bernstein2001). As opposed to the classic approach and the prototype model, features play a subordinate role because they do not define the concept, but instead constitute the product of our underlying knowledge. This hypothesis informs the existence of a non-conventional type of categories that are referred to as ad hoc: ‘an ad hoc category is a novel category constructed spontaneously to achieve a goal relevant in the current situation (e.g., constructing tourist activities to perform in Beijing while planning a vacation)’ (Barsalou, Reference Barsalou and Hogan2010, p. 86).
As these categories are created specifically for a given and particular purpose at the time of use, exemplars of an ad hoc category do not even have to share common features, although they have a hierarchical structure, with more or less representative exemplars, the same as natural categories (Croft & Cruse, Reference Croft and Cruse2008, p. 129). The consideration of belonging stems from the fact the member fulfils an objective or ideal (Best, Reference Best2002, p. 407; Kövecses, Reference Kövecses2006, p. 27; Sawyer, Reference Sawyer2007, p. 120): for example, the members of the category foods eaten when on a diet will be judged according to the objective of being low in calories – the ideal is to have none at all – and this will determine their belonging to the category.
This internal structure is generally as stable and robust as that of familiar taxonomic categories (Cassol Soro & Boto Ferreira, Reference Cassol Soro and Boto Ferreira2017; Mauri, Reference Mauri, Blochowiak, Grisot, Durrleman and Laenzlinger2017). Nevertheless, in contrast to these, ad hoc categories are not embedded in the memory and not possible without a context (Barsalou, Reference Barsalou and Hogan2010, p. 86). Nonetheless, the repeated use of an ad hoc category renders it highly familiar and consolidates it in the memory, constituting what Barsalou (Reference Barsalou and Hogan2010, Reference Barsalou, Fiorentini, Mauri and Goria2021) calls a goal derived category. They both have a bottom-up structure, with the category being generated by the exemplar that meets the expectations held towards the category, and not vice versa (Barsalou, Reference Barsalou, Fiorentini, Mauri and Goria2021; Mauri, Reference Mauri, Blochowiak, Grisot, Durrleman and Laenzlinger2017).
This study proposes Objects laid on the table for a meal (hereinafter Objects laid on the table or simply Table) as an example of this type of category. The aim of facilitating a meal allows relating such disparate items as bread, fork, tablecloth and vase, and the most typical object will be considered the one that best suits that purpose. We therefore refer to this as an ad hoc category, although strictly speaking it is a goal derived category, understanding it to be embedded in the long-term memory because of the daily chore of laying the table.
Like Barsalou, Lakoff shows the ways in which contextual knowledge wields its influence on the formation of categories, assuming that the lists of features are insufficient for fully representing a concept: concepts are constructed on the basis of cultural criteria that may or may not be immediately obvious (Berko & Bernstein, Reference Berko and Bernstein2001). Accordingly, the radial categories proposed by Lakoff (Reference Lakoff1987) – or complex categories according to Langacker (Reference Langacker and Rudzka-Ostyn1988, pp. 134–135) – are also framed within the knowledge approach and are based on the aforementioned notion of family similarity. A word activates multiple cognitive domains, which in turn activate other conceptual networks, which need not be directly related to the former. These categories have a prototypical nature with a central member and a network of links to other members (Brugman & Lakoff, Reference Brugman, Lakoff and Geeraerts2006, p. 109). Nevertheless, ‘both central and noncentral subcategories have their own representations, and no properties of subcategories can be predicated from the central subcategory’ (Lewandowska-Tomaszczyk, Reference Lewandowska-Tomaszczyk, Geeraerts and Cuyckens2007, p. 148).
As an example of a radial category, we have chosen Games and pastimes, consisting of sundry subcategories (e.g., card games, board games, children’s games and sports), some more central than others, which do not have to be directly related to each other, as noted by Wittgenstein (Reference Wittgenstein, Laurence and Margolis1999, p. 171): ‘Consider for example the proceedings that we call “games". I mean board-games, card-games, ball-games, Olympic-games, and so on. What is common to them all? […] For if you look at them you will not see something that is common to all, but similarities, relationships, and a whole series of them at that’.
Finally, the category Countryside has a schema organization, specifically, corresponding to what De Vega (Reference De Vega1995, p. 393) calls visual schema, or to the scenes according to Mandler (Reference Mandler1984, pp. 77–78): ‘scene knowledge consists of inventory and spatial relation information that allows us to recognize a place as a living room or a playground’. Schemas are knowledge structures that represent stereotypical situations that are frequently encountered. There are acquired based on prior experience and serve as interpretative models of the new situations we face (Cifuentes Honrubia, Reference Cifuentes Honrubia1994).
Indeed, our memory stores incomplete schema consisting of fixed and variable parts (Minsky, Reference Minsky and Winston1975; Rumelhart & Ortony, Reference Rumelhart, Ortony, Anderson, Spiro and Montague1977; Rumelhart, Reference Rumelhart, Spiro, Bruce and Brewer1980 apud Téllez, Reference Téllez2005, pp. 95–97). Once the schema has been activated, the slots in it can be filled with specific information on the real or specific event described, or they can be completed with the default value typically associated with the framework in question. This default value is used when the outside world does not provide any other value and may generate false memories (Wilson & Keil, Reference Wilson and Keil2001, p. 729).
Schemas also represent knowledge at all levels of abstraction; they incorporate both episodic and semantic declarative knowledge, they are governed by an affective attitude that is reflected in memory processes and are arranged into hierarchies: schemas include others of a more elementary nature and, in turn, constitute the sub-schema of others (Crespo León, Reference Crespo León1997, p. 277; Téllez, Reference Téllez2005, p. 97).
Countryside therefore corresponds to a very general schema, which includes other more specific scenes according to their hierarchical arrangement. The examples reviewed for this category are numerous and diverse, depending on the sub-schema chosen (village, square, farm) and may refer to agents (old people, shepherd, farmer, labourers), objects (animals, plants and geographical features), activities (rest, harvest) or feelings and sensations (tranquillity, happiness), among others directly related to the proposed schema or to one of its sub-schemas.
3. Exploring the types of categories according to network analysis
The main aim here is to corroborate the descriptions proposed by linguistic theory for characterizing the different types of categories (see Section 2.2) through network analysis that underpins their lexical output in a verbal fluency task. Our general hypothesis is that different categorial models condition the architecture of the network they generate, as well as the way in which the speakers access and move within it.
The verification of this general hypothesis involves two levels of analysis: (1) mesoscopic (exploring each network’s structure, as well as its access and navigability, and (2) microscopic (analysing the relations established from the central nodes and paths that traverse them). An original aspect in this kind of studies is that both levels of analysis consider the way in which informants enter and explore the network, not only as a construct of specific structural characteristics but also as a process whose orchestration involves the speakers themselves. In turn, at each level of analysis, the general hypothesis is broken down into other more specific ones, as shown in Tables 1 and 2.
Table 1. Hypotheses at mesoscopic level

Table 2. Hypotheses at microscopic level

4. Methodology
4.1. Semantic fluency task and network construction
Participants were 680 preuniversity students (17 to 18 years old), 290 males and 390 females, all of whom are native speakers of the Spanish spoken in central Spain. They completed a semantic fluency task during a class period in their scholar institutions. The task adopts the methodological parameters from lexical availability studies in linguistics, as part of the Pan-Hispanic Lexical Availability Project (López Morales, Reference López Morales and Catalán2014), with written tests conducted for 16 categories and two minutes for each. Four prompts were chosen from the entire set.: Animals, Objects laid on the table for a meal, Games and pastimes and Countryside. During the processing of transcribing, the responses were corrected and edited according to the most recent normative orthography of Spanish (ASALE & RAE, 2010). Morphological variables were standardized under a single lemma-type (gender and number, e.g., maestro, maestra, maestros and maestras into maestro ‘teacher’) and duplicate words were removed (see Hernández Muñoz (Reference Hernández Muñoz2006) for a more comprehensive description of the data editing processes). The project encompassing this research received approval from the Ethics Committee for Research at the University of Salamanca.
Data processing and network construction have been performed with the tool LexPro (Hernández Muñoz et al., Reference Hernández Muñoz, Tomé, López García, Díaz and Muñoz2025), which directly provides the calculations specified in 5.1 for the general characterization of the categories according to their results in the semantic fluency task: total number of words (tokens), the number of different words or terms (types), and the average per informant (see Table 5). LexPro also caters for the generation of different kinds of graphs and provides traditional mathematical indices for the description of nodes and networks.
The created graphs consist of undirected bigrams, representing all pairs of words produced consecutively by the informants without considering the direction in the production chain, following traditional models of network creation based on experimental data from semantic fluency tasks (Feng & Liu, Reference Feng and Liu2023; Goñi et al., Reference Goñi, Arrondo, Sepulcre, Martincorena, de Mendizábal, Corominas-Murtra, Bejarano, Ardanza-Trevijano, Peraita, Wall and Villoslada2011; Zemla & Austerweil, Reference Zemla and Austerweil2018). To generate these graphs, the words provided by a single informant were excluded, and edges with a weight of one were removed (Borodkin et al., Reference Borodkin, Kenett, Faust and Mashal2016). The mathematical measures considered for the analysis are described in the next section. The SPSS 26 package has been used for the complementary statistical analysis.
4.2. Measures used for the analysis
The verification of each hypothesis formulated in Section 3 has involved the use of different mathematical indices. They are all described in Tables 3 and 4.
Table 3. Measures considered for each hypothesis in the mesoscopic analysis

Table 4. Measures considered for each hypothesis in the microscopic analysis

5. Analysis
5.1. General data on categories
Table 5 presents the general results of the semantic fluency test. Animals is the most productive category, with 14721 tokens and an average of 21.65 words per informant, compared to Objects laid on the table, which records a lower average. The analysis of variance confirms that the differences observed in the number of words per informant in each category are statistically significant (F[3.2716] = 583.003, p = 0.000). Post-hoc comparisons used the Scheffé test indicate that the differences between all possible category pairs are significant (p = 0.000).
Table 5. Quantitative indices of the semantic fluency task

5.2. Mesoscopic level
5.2.1. Network size
‘The traditional measures of Average Path Length and Diameter are closely similar in Animals and Objects laid on the table, with a slightly broader network for Games and Countryside’. Considered in isolation, there is no difference between our lexical sets. With a view to increasing the explanatory power of our analysis, we have chosen to consider the number of communities they include (Table 6) and the number of nodes in each community (see Supplementary Materials). This property is assessed as a continuum: in itself, the number of nodes is not so relevant as the degree of homogeneity or heterogeneity of the community’s architecture.
Table 6. Communities in each category’s networks

Furthermore, as general properties of the four networks, we can confirm that they all have a small-world architecture (sigma > 1) and a certain degree of modularity (see Section 5.2.2.), whereby their structure responds to the required models of a mental lexicon network. As we shall see in due course, the greatest differences among the four categories are to be found in the topology of their modularity.
Animals has nine communities, six of which (66.7%) exceed the average of nodes per community, with over 30 members (between 51 and 32). Objects laid on the table has 15 communities, with 40% exceeding the average (10.87 nodes), and only the first one has more than 30 members. Games has 19 communities, with seven of them exceeding the average (36.8%), with between 50 and 22 members. Finally, out of the 31 communities in Countryside, only eight exceed the average (25.8%). This category has three very large communities (87, 77 and 58 members), seven with between 35 and 12 members and 22 with few members (15 of which have only two members).
These results indicate that, in agreement with H1, the network corresponding to the natural category is denser than all the others. Animals has fewer communities and they are more uniform in terms of the number of their members. At the other end of the scale, Countryside, has numerous communities and is highly diverse in terms of the number of nodes it contains. Like Countryside, Games creates a large network, although its communities are less heterogeneous. Both Games and Objects laid on the tables are located between these two extremes (i.e., Animals and Countryside), in the middle of this continuum. Finding the differences between them requires exploring the characteristics of their communities.
5.2.2. Internal and external relations of network communities
H2 refers to the interconnection between the members of a community and the links they maintain with other communities in the network. Section 4.2 has already explained that these relations are measured by calculating the EXTINT ratio, where a value close to 1 indicates that the connections between the members of a community are similar to those established towards the outside of the community. This measure, therefore, is determined by the inward or outward degree, and needs to be interpreted together with modularity, which refers to how network nodes are grouped into communities and how these communities are defined. A network is closely interconnected when it has a high EXTINT ratio and a high degree of modularity.
According to the initial hypotheses, as Table 7 shows, the community of Animals is interconnected more towards the outside than towards the inside (EXTINT ratio of 0.63) than those of Games and Countryside (both with an EXTINT ratio of 0.43). Although Objects laid on the table records the highest EXTINT ratio (0.95), this figure should be interpreted in relation to the lack of modularity (0.29). In other words, as opposed to Animals (modularity of 0.45) and Games and Countryside (both with a modularity of 0.52), Objects laid on the table is not arranged into defined communities. The nodes tend to be connected with all the other nodes in the network (both those in its community and those in others) because there is no clearly defined structure, which is consistent with this category’s ad hoc nature.
Table 7. Measures of association in each network’s communities

A close study of the external and internal degree reveals that Animals and Objects laid on the table have more edges towards the outside than towards the inside (1686 internal ones/2352 external ones, and 600 internal ones/1030 external ones, respectively), whereas the links towards the inside in Games and Countryside are more numerous (1122 internal ones/812 external ones, and 1430 internal ones /862 external ones, respectively). Nevertheless, the average for internal strength is higher in all cases.
5.2.3. Existence of a central community
H3 considers that the different nature of the categories impacts upon the existence of a central subcategory that the participants access above all, as the lexical set evoked the most. This central community has been located by calculating the PCom, which provides the percentage of participants that mention the members of each subcategory (see Supplementary Materials for members visited the most in these communities).
In Animals, over 50% of the participants account for the first five terms corresponding to community 1, which contains the following wild animals: lion, bear, elephant, tiger, monkey, giraffe, crocodile, zebra, hippopotamus and panther. The first two (lion and bear) are shared by more than 80% of the informants.
In Objects laid on the table, C3 is the central subcategory, but only three of its members are evoked by more than 50% of the informants. The common denominator among the nodes in this community is that they are part of a dinner service: glass, plate (both named by more than 80% of the sample), jug, bottle, spoon, wine-glass, teaspoon, serving dish, fork and knife.
In Games, the community that is most evoked includes sports: football, basketball, tennis, volleyball, handball, swimming, table-tennis, bowls, rugby and hockey. Nevertheless, only the two first members of the community are evoked by more than 50% of the informants, and none exceeds an 80% mention.
Countryside is the most accessible community, consisting of nouns that refer to natural features (tree, animals, path, flower, stone, bird, grass, earth, insect and nature), which is consistent with the weight of the visual component in this schema category. In this case, four members of the central community are evoked by more than 50% of the informants, with tree and animals being mentioned by more than 80%.
It is noteworthy that although the communities have been formed using mathematical models, a qualitative analysis identifies shared semantic features, whereby the structural strength is backed by semantics.
According to the initial hypothesis, both Animals (93%) and Countryside (92%) have a central subcategory that the speakers refer to most in the semantic fluency task, and in Animals especially, the percentage of visits to the nodes remains very high for almost all the members of that central community. Objects laid on the table also has a dominant community (99%), although its exemplars cease to record a high percentage sooner than in Animals. In the case of Games, however, no community receives so many visits (70% with a sharp drop as from the third exemplar), whereby it may be posited that this radial category does not have a clearly central community. Thus, in contrast to the data in Section 5.2.2, they distinguish the radial category Games from the schema Countryside.
5.2.4. Lexical access and negotiating the network
The final hypothesis at mesoscopic level (H4) refers to the part of the network that each informant traverses. In the more closely interconnected category, with internal and external relations between communities, the speakers may move around more within them, whereby each informant visits a higher percentage of their constituent nodes. By contrast, a participant in the more disperse and heterogenous categories may be confined either to a single community or to a reduced number of nodes.
With a view to analysing the participants’ behaviour, we may consider the measure of the percentage of the network (PNet) the participants visit in each cluster (Table 8). The participants in Animals and Objects laid on the table visit around twice the amount of the network compared to Games and Countryside. In other words, each participant visits a much more extensive part of the network during the task. In turn, the histogram (Figure 1) presents the PNet data: the y-axis reports the number of participants and the x-axis the percentage of the network each one of them visits. Animals records the highest percentage of the network visited (with a highly representative figure of 100 participants negotiating 8% of the network compared to the peaks for Countryside of 60 informants visiting barely 2.5%). These data depict a continuum from the natural category through to the schema: Animals is the category in which an informant visits a greater proportion of the network, followed closely by Objects laid on the table, Games and, finally, Countryside.
Table 8. Gaussian curve averages for each category in the histogram


Figure 1. PNet histogram.
5.3. Microscopic level
5.3.1. Relations between the central node and other nodes
Table 9 presents the measures considered for verifying H5, regarding the relations between the central node and all the other nodes in each network. The significance of the most central word is evaluated through the values corresponding to the three strongest nodes (-2) in each category, and an asterisk indicates those nodes considered to be central. This is not the only way of measuring a node’s significance within the overall network, as other relevant values are frequency (indirectly represented in strength), age of acquisition (Steyvers & Tenenbaum, Reference Steyvers and Tenenbaum2005) and lexical availability (Hernández Muñoz & Tomé, Reference Hernández Muñoz, Tomé and del Barrio2017). All these mathematically based values could represent aspects of the prototype, although the conceptualization of the last one of these is based on cognitive attributes that need not necessarily correspond to a network’s centrality, especially in more delayed approaches to abstraction (Geeraerts, Reference Geeraerts and Geeraerts2006).
Table 9. Measures of association for the three strongest nodes in each category

Objects laid on the table creates the network with the most salient node in terms of centrality and the strongest connections. Tablecloth is the object that best fulfils the contextual category’s purpose (centrality 0.41). Nevertheless, the hypothesis is not fully confirmed in the case of Animals: centrality is shared among several members of the central community, all of which are visited very frequently (the three indices considered are similar). Nonetheless, these two categories with a more taxonomic structure have central nodes with a higher number of overall strength (over 0.2) than Games and Countryside, where the concept of central node is diluted (below 0.2). It is still significant, however, that tree is the most central category in this last category, as it has high visual salience, and at the same appears as a hyponym in animals, which requires the abstraction of specific features and gives rise to one of the communities that are part of the schema’s framework.
5.3.2. Paths traversing the central node
H6 addresses the likelihood that a short path between two nodes traverses what is considered to be the central node. As noted, this characteristic is measured by the indicator betweenness centrality (BC) (Table 10).
Table 10. Betweenness centrality for the central nodes in each category

According to these data, the central node that is traversed by more short paths is the ad hoc category Objects laid on the table, followed by the schema Countryside, the radial Games and, finally, by Animals. Speakers in the ad hoc category return more often to the item that best fulfils the category’s purpose (a tablecloth is the first object laid on a table before a meal, and which therefore orchestrates its construction. Contrary to the initial hypotheses, the informants in the natural category tend not to traverse the central node, as accessing a very dense network, with clearly defined communities (see Section 5.2) and several nodes with a high degree of centrality (see Section 5.3.1), speakers have numerous possible paths available to them, which means they do not need to traverse the same node.
6. Discussion
This study’s overriding aim was to verify whether the measures arising from networks based on empirical data supported the characteristics that linguistics has attributed to the different types of categories that structure the semantic memory. Our general hypothesis was that an analysis of the network would enable us to identify its nature, not only regarding the network’s topology but also through a speaker-based linguistic approach, involving how speakers access semantic knowledge, recover it and produce it. The network framework and mathematical techniques used allow us to consider both the structure and the process of storage and interaction of lexical items; in other words, understanding both the structure and the dynamics of a lexical system in an approach that does not separate language from the human mind (Beckage & Colunga, Reference Beckage, Colunga, Mehler, Lücking, Banisch, Blanchard and Job2016). This study uses the term network dynamics as the prior process through which the speaker generates the network’s very structure; all within an approach of linguistics based on speakers. This view also aligns with theories that conceive categorization from a processual perspective, rejecting the classical view of categories as completely fixed structures (Barsalou, Reference Barsalou and Hogan2010).
The mesoscopic and microscope levels have proven to be suitable for this analysis. As our data are the result of an experimental approach and the speakers themselves are the ones that have provided the data for the analysis, we understand that the actual process of categorizing reality responds to active generation in the use of language, with users extracting and creating the network at this mesoscopic level by contributing a differential perspective on the linguistic networks analysed. This reinforces the idea that categories are not independent of the experiential context, in line with the proposals of cognitive semantics (Lemmens, Reference Lemmens and Riemer2015). The measures of analysis, therefore, have also considered the process of moving around in the network during the lexical recovery. Our objective has been achieved by supplementing traditional measures with others created specifically for this purpose, which may be used by ensuing studies focusing on the mesoscopic and microscopic levels of all kinds of linguistic networks. The EXTINT ratio provides information on the topography of communities, while PCom, PNet and BC describe the network as a process through which the speakers in the category fluency task construct the network on the basis of relations between words that guide their production, while at the same time, they move through the cluster of terms available to them in the dynamic system of their mental lexicon. Thus, an exemplar like cow is retrieved in the category Animals as a typical representative of the subcategory Farm Animals, but it also emerges in relation to the schema Countryside, guided by the activation of a related subschema like Farm, which encompasses terms with which it does not share semantic features but does share a spatial context, such as tractor, barn or farmer.
In general, although all the conceptual categories studied are small-world networks and have a certain degree of modularity, it is precisely this latter aspect that gives rise to their complexity and the differences in their structure, to the extent that, as we see in the specific hypotheses, the categories are defined not so much by the properties of their nodes as by the communities forming them.
There now follows a review of each one of the specific hypotheses considered. The results on the network’s structure and architecture confirm H1: there is a gradation in the number of communities and their members from the natural category, with fewer communities consisting of a more numerous cluster of members, through to the schema category that has a more disperse structure and fewer member because it is informed by the broad experience of the environment it describes (Cifuentes Honrubia, Reference Cifuentes Honrubia1994; Mandler, Reference Mandler1984). Nevertheless, the ad hoc and radial categories occupy more central positions on the continuum and require other supplementary measures for differentiating them.
H2 describes the nature of communities’ internal and external relations. Our predictions are once again upheld, as Animals is the category with the highest EXTINT ratio and the greatest degree of modularity. So, too, the natural category, informed by the cumulative experience of speakers from an early age, which reinforces a hierarchical taxonomic structure with highly structured and closely clustered subcategories with the base terms being firmly embedded in shared knowledge (Rosch, Reference Rosch1975). It should be noted that the ad hoc category has the highest EXTINT ratio, which indicates that the nodes have a high level of connections almost in equilibrium both inside and outside the communities, albeit with reduced modularity. This is due to the structure of a contextual category created around the exemplars that best fulfil the category’s original purpose (Barsalou, Reference Barsalou and Hogan2010, Reference Barsalou, Fiorentini, Mauri and Goria2021), whereby all the categories gravitate around it or other nodes and do not lead to independent clusters. To conclude, H2 does not differentiate here between the radial and schema categories, as they both record similar data.
H3 focuses on the existence of a central community identified by the percentage of participants visiting it. This figure describes the network as a process and incorporates a user’s movements. The results corroborate our hypothesis, as both the taxonomic and the ad hoc categories have a central community that is frequently visited, responding to a highly stable and robust structure (Mauri, Reference Mauri, Blochowiak, Grisot, Durrleman and Laenzlinger2017), which is prototypical in the former, and created in the latter on the basis of an exemplar (or set of exemplars) that fulfils its purpose. There is also a central category in the schema, which could be the community with the nodes that represent the most salient entities for the senses, visited by 92% of the speakers. Nevertheless, there is no central community in the radial category because of its architecture based on the adjacency of relatively isolated subcategories, with features that are not necessarily shared by all of them (Wittgenstein, Reference Wittgenstein, Laurence and Margolis1999). This is the index that best defines the radial structure compared to all other categories studied.
The final hypothesis at mesoscopic level addresses the percentages of participants’ movement around the networks. The results reveal that the most travelled categories are the natural and ad hoc ones, recording a PNet that is approximately twice that of the radial and schema ones. This reinforces the findings in H2 and H3, where the natural and ad hoc categories presented a denser and more stable network model (with higher EXTINT ratios): communities that are more densely interconnected not only towards the inside but also towards the outside allow traversing more nodes and subcategories within the network, with numerous shortcuts between the different lexical areas. This ease of communication within the network is also enhanced by a fewer number of communities, especially in the natural category (H1). The dispersion of the radial category (H1) and the scant connectivity of the ad hoc and radial ones (H2) decrease the likelihood of speakers accessing and traversing an extensive part of the network.
There now follows an analysis of the microscopic level. As noted in the theoretical section, the theory of categorization in cognitive psychology develops in step with the consideration of the existence of a prototype that draws together all the other exemplars in the category, blurring its boundaries (Mervis & Rosch, Reference Rosch and Mervis1975). Nevertheless, the first version of the prototype was being abstracted in the theory’s development until a prototypical categorization was attained as a mechanism for the organization of knowledge, albeit not necessarily an identifiable prototype (Geeraerts, Reference Geeraerts and Geeraerts2006). This reinforced typicality as a decisive cognitive factor in the categorial structure that also affects language processing (Hernández Muñoz et al., Reference Hernández Muñoz, Izura and Ellis2006). This evolution in the concept of prototype towards abstraction and the non-existence of a single central exemplar is precisely what we have found in the results of our model of taxonomic category Animals. There is no single central node, but rather a central community in which several nodes share the same degree of centrality and associative force (lion, elephant and bear). The figure of the central node is much more readily apparent in the ad hoc category. Tablecloth is the word that most closely fulfils the objectives of the conceptual cluster, constructed bottom-up from the exemplar (Barsalou, Reference Barsalou, Fiorentini, Mauri and Goria2021). In the radial and schema categories, the articulating role of a node’s centrality is diminished, with the most cohesive nodes within the network recording a lower degree of centrality and less overall strength than in the natural and ad hoc ones (football and tree, respectively). Furthermore, tree has the highest degree of visual salience in a landscape, hence the reason it tends to articulate all the other nodes in the network, thereby reinforcing the type of experiential or sensory-motor based semantic knowledge (Vigliocco, Meteyard, Andrews, & Kousta, Reference Vigliocco, Meteyard, Andrews and Kousta2009). Nevertheless, the relations between the concept of prototype and the different measures of centrality of the semantic network are complex and require a more detailed exploration.
H6 refers to the number of paths that traverse the most central element. Once again, the prediction is fulfilled regarding the ad hoc category, given that tablecloth is the central node that is traversed by the shortest paths, which is consistent with the central exemplar’s theoretical significance in this mechanism for structuring the lexicon. However, as one of the least expected results, the hypothesis is not fulfilled in the natural category, perhaps because of the outcome encountered in H5, whereby there is no single central element, but instead a cluster of central nodes, members of the network’s most representative community. Again, these results reveal an arrangement of the taxonomic categories that does not have a single central prototype, but instead a cluster of exemplars that constitute a central community.
Taking the mesoscopic and microscopic measures studied here in their entirety, we may affirm that all the categories have recorded more salient indices that define their construction process. The natural or taxonomic category is well established in semantic knowledge, with few yet broad categories that are interconnected and stable (high EXTINT ratio and high modularity), with a cluster of central nodes and subcategories that carry a great deal of weight in the movement within them (high PNet). One might refer to a prototypical community whose members occupy similar positions of centrality and are closely linked to all the others, where the notion of a central member fades. As regards the ad hoc category, the bottom-up structure is ratified in all the hypotheses around a community (mesoscopic) and an exemplar (microscopic) in which all the other nodes are orchestrated, as categorial membership is based on the fulfilment of its objective (high node centrality and the central community’s high PCom). The radial community upholds its definition around a sum or adjacency of subcategories with a certain degree of independence. The feature that sets it apart from the others is its PCom, which reveals the absence of a central category. Finally, the schema category records the greatest dispersion, with numerous communities that are formally heterogeneous. Nevertheless, a crucial aspect is the existence of a central community and node in which perceptual salience prevails.
7. Conclusions
An overarching aim in studies on language and cognition has involved the use of interdisciplinary methods to show whether or not theoretical linguistic models have their correlation in speakers’ language output. The analysis of sets of linguistic data, with such highlights as the empirical data from psycholinguistic experimental tasks administered both online and offline, caters for this kind of study. Although data from behavioural tasks have been used to test sundry categorial hypotheses, our methodological approach is original because it includes a mathematical analysis derived from complex network theory as interaction systems. Our study therefore reviews the theoretical descriptions of four types of conceptual categories represented in the semantic memory based on semantic networks constructed with data gathered from a category fluency task. The results confirm the linguistic proposals and qualify certain aspects of the cognitive description, reinforcing the conception of categorization as a dynamic and adaptive process, in which different types of information are integrated.
Within this context, the mesoscopic and microscopic dimensions emerge as the most suitable ones for analysing semantic categorization. Understanding the network of specific subsections of lexicon ⎯ especially when they are of a different nature ⎯ enables us to further explore the features that arrange and construct the overall network. One of the conclusions reached by this analysis is that the description of the categories requires the use of a broad set of mathematical measures, as most of them are continuums, recording varying degrees of internal arrangement and, in isolation, they do not allow for absolute assessments to be made of the network’s behaviour. There is, moreover, a need to consider metrics for relating the structure to the informant’s dynamics within the networks (the network as a process created by the speaker when extracting the semantic memory).
A future extension of this study could include more categories for each one of the types studied to verify whether they all fulfil the parameters yielded by the data presented here. In turn, although the study’s high number of adult participants is an assurance of relatively stable tendencies, it might also be necessary to extrapolate it to other populations. As regards methodological limitations, it might be worth exploring not only direct relationships between words (bigrams) but also indirect ones (linear graphs), which may improve the evaluation of the relatedness of words, specifically when considering the qualitative nature of associations.
As a final consideration, structural features of networks account for certain psycholinguistic phenomena observed in tasks requiring lexical and semantic processing. For instance, the degree or local clustering coefficient of a node may influence on how easily a spoken or written word is recognized, new words are learned, lexical items are recalled in memory tasks (Baronchelli et al., Reference Baronchelli, Ferrer-i-Cancho, Pastor-Satorras, Chater and Christiansen2013; Siew et al., Reference Siew, Wulff, Beckage and Kenett2019), or how phonological forms are processed in a variety of tasks (Vitevitch et al., Reference Vitevitch, Goldstein, Siew and Castro2014). Conversely, psycholinguistic tasks have been employed to validate networks as plausible models of memory retrieval. For example, semantic similarity judgments made by speakers have demonstrated the psychological relevance of these networks (Zemla & Austerweil, Reference Zemla and Austerweil2018). Thus, the relationship between networks and language processing is profoundly interconnected. Therefore, our proposal may have significant psycholinguistic implications for future research. Variations in speaker behaviour during lexical retrieval ⎯ reflected in topology and dynamics of networks ⎯ are category-oriented, making it crucial for understanding lexical operations, such as lexicon learning, or methodological considerations, such as selecting appropriate stimuli for verbal fluency assessments.
It is essential to observe linguistic processes from interdisciplinary perspectives to discover innovative tools that pave the way for reflective processes that have hitherto not been explored through traditional pathways. The enrichment between disciplines is multidirectional, as our study shows. Complex network theory may thus benefit from the search for new patterns of movement through a network based on speakers and the linguistic theory of nuances and reviews of prior models. Only in this way will we be able to take small steps towards addressing the vast complexity of the linguistic phenomenon.
Supplementary material
The supplementary material for this article can be found at http://doi.org/10.1017/langcog.2025.10.
Data availability statement
The data that support the findings of this study are openly available in OFS at https://osf.io/xjyr9/?view_only=cc992e6939694b97823fd8d19be9831e.
Acknowledgements
This paper is supported by the project ‘DispoGram. Disponibilidad gramatical del español’, PID2020-120436GB-I00, funded by Ministerio de Ciencia e Innovación del Gobierno de España.